---
language: en
license: apache-2.0
tags:
- text-classification
- intent-classification
- onnx
- quantized
- distilbert
datasets:
- custom
metrics:
- accuracy
model-index:
- name: distilbert-intent-classifier-onnx-int8
  results:
  - task:
      type: text-classification
      name: Intent Classification
    metrics:
    - type: accuracy
      value: 0.85
      name: Accuracy
      verified: false
---

# DistilBERT Intent Classifier (ONNX INT8 Quantized)

## Model Description

This is an **INT8 quantized ONNX** version of a fine-tuned DistilBERT model for 6-class intent classification.
Optimized for **browser-based inference** with ONNX Runtime Web.

### Intent Categories

1. **arithmetic** - Simple calculations, unit conversions, percentage math
2. **symbolic_reasoning** - Mathematical proofs, physics derivations, complex logic
3. **factual_lookup** - Search-engine style questions (history, geography, etc.)
4. **creative_synthesis** - Fiction writing, poetry, roleplay, brainstorming
5. **code_generation** - Writing, refactoring, or auditing code
6. **security_risk** - PII leaks, API keys, injection attempts

## Model Details

- **Base Model**: distilbert-base-uncased
- **Quantization**: INT8 Dynamic Quantization (AVX512-VNNI)
- **Model Size**: 65.15 MB (75% smaller load time vs. unquantized)
- **Format**: ONNX
- **Target Platform**: Browser (ONNX Runtime Web) / CPU inference

## Performance

- **Load Time**: ~75% faster than unquantized PyTorch model
- **Inference Speed**: 2-3x faster on CPU
- **Accuracy**: >85% on arithmetic and search (factual_lookup) intents

## Usage

### Browser (ONNX Runtime Web)

```javascript
import * as ort from 'onnxruntime-web';

// Load model
const session = await ort.InferenceSession.create('model.onnx');

// Tokenize input (you'll need to implement tokenization in JS)
const inputIds = tokenize("What is 25 + 17?"); // Returns Int64Array

// Run inference
const feeds = {
  input_ids: new ort.Tensor('int64', inputIds, [1, inputIds.length]),
  attention_mask: new ort.Tensor('int64', attentionMask, [1, attentionMask.length])
};

const results = await session.run(feeds);
const logits = results.logits.data;
```

### Python (Optimum)

```python
from optimum.onnxruntime import ORTModelForSequenceClassification
from transformers import AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("RunsOnBacon/distilbert-intent-classifier-onnx-int8")
model = ORTModelForSequenceClassification.from_pretrained(
    "RunsOnBacon/distilbert-intent-classifier-onnx-int8",
    provider="CPUExecutionProvider"
)

inputs = tokenizer("What is 25 + 17?", return_tensors="pt")
outputs = model(**inputs)
predicted_class = outputs.logits.argmax(-1).item()
```

## Training Data

- **Dataset**: 1,200 synthetic examples (200 per category)
- **Special Feature**: 30% of arithmetic and factual_lookup examples are "cloaked" with sophisticated prompt structures to teach robustness
- **Split**: 80/20 train/validation

## Limitations

- Trained on synthetic data - may not generalize to all real-world variations
- Optimized for English language only
- Security risk detection is pattern-based, not comprehensive

## Model Card Contact

- **Repository**: https://huggingface.co/RunsOnBacon/distilbert-intent-classifier-onnx-int8
- **Original PyTorch Model**: https://huggingface.co/RunsOnBacon/distilbert-intent-prompt-classifier